A working definition, since I couldn't find one anywhere.
The term "AI-native" is starting to show up in our profession. AI-native firms, AI-native accountants, AI-native workflows. I have yet to see anyone actually define it. Mostly it gets used to mean "uses AI a lot," and I don't think that's good enough, because the profession is about to sort itself by this word and there should probably be a shared consensus as to what it means first.
So here's my attempt. I'm not claiming any authority here, though I've been describing my own firm as AI-native since I wrote about it for GSCPA's Current Accountants this spring, and it's past time I defined what I actually mean by it. I spent the past year inside a startup building AI agents that turn hard tax problems into expert-level advice, I run my own practice on AI tooling today, and I've spent the past few months talking with accountants and firm owners about how they're using it. That's the seat I'm writing from.
Why "native" is the right word
We've been through this once before, with the cloud.
Fifteen years ago, plenty of firms "used the cloud." They had file sync and a hosted app or two, bolted onto a practice that worked exactly the way it had always worked. A cloud-native firm was a different animal. It assumed the cloud from day one, and the practice was shaped by what that assumption made possible: work from anywhere, clients from anywhere, a tech stack instead of a server closet.
Same word, two completely different firms. The difference was never how much cloud software they bought. The difference was whether the assumption came first.
Here's the same idea as a lineage. Most firms today live somewhere in the middle column.
| The traditional firm | The tech-enabled firm | The AI-native firm | |
|---|---|---|---|
| The work | Recording what happened | Running software that records what happened | Reviewing AI's first pass and deciding what happens next |
| The rhythm | Batch work and a monthly close | Bank feeds and a faster close | Continuous. Work gets reviewed as it lands, not at month end |
| The data | Manual entry and formulas | Exports and pivot tables | Plain-English questions, whole-ledger answers |
| What clients pay for | Accuracy | Speed | Judgment |
That bottom row is the whole story.
So here's the definition I'd offer:
An AI-native CPA runs a practice designed around the assumption that AI does the production work: the doing is delegated, the judgment is the product, and the practice can build what it can't buy.
The five marks
Definitions are cheap, so here's how I'd actually recognize one.
1. AI does the first pass. They own the last one.
In an AI-native practice, first drafts are assumed. The research memo, the client email, the workpaper tie-out, the first pass of the return. In my own practice it also looks like asking the ledger a question in plain English and getting an answer with receipts, instead of building the pivot table that would find it. The CPA's time concentrates at the review layer, because that's where the license lives and that's what the client is actually buying. And to be clear about what this doesn't mean: nothing reaches a client unreviewed. The AI-native practices I've seen are stricter about review than traditional ones, because they designed the gate on purpose instead of inheriting it from how things have always been done.
2. They build what they can't buy.
When we surveyed 437 accounting professionals for The State of AI in Accounting Firms, 18% were already building custom workflows and tools. A script that renames and files source documents. A skill that drafts the deliverable in the firm's own format. An internal app the market simply doesn't sell. Five years ago that was absurd for a small firm. Today it's a weekend. The AI-native CPA treats "the tool doesn't exist" as a to-do item, not a dead end.
3. They price judgment, not keystrokes.
When production time collapses, hourly billing quietly becomes a penalty for competence. The AI-native CPA has already made peace with this: the client was never paying for keystrokes, they were paying for the judgment that knows whether the answer is right and the license that stands behind it. Price follows the value of the answer, not the hours of the doing. This is the mark that takes the longest to adopt, because it's the one that lives in your head instead of your tech stack.
4. Their permission is in writing.
This is the unglamorous one, and I believe it's the one that separates firms that move from firms that stall. A current WISP that names the tools. An AI policy. A considered position on when ยง7216 consent applies. Not because the paperwork is the point, but because a firm that has written down what it's allowed to do gets to move fast without looking over its shoulder. I keep meeting capable people who are stalled for no technical reason at all. They're waiting for someone to tell them they're allowed to start. AI-native firms gave themselves the answer, in writing.
5. They stay in the work.
This one I hold personally. I don't think I can credibly help firms adopt AI without doing client work myself, and I'd apply the same test to the definition: the moment you stop doing the work, you start describing other people's. Being AI-native doesn't mean graduating out of the profession into commentary. It means staying in the reps while the reps change.
Where the profession actually is
The same survey puts the engaged segment of the profession on a ladder. 45% are still dabbling with their main AI assistant. 32% use it daily. 18% are building their own workflows and tools.
Daily use alone doesn't make a practice AI-native. I'd describe the ladder as AI-curious (trying tools), AI-assisted (recurring workflows you actually trust), and AI-native (the practice is designed around the assumption). Most of the profession's engaged half sits somewhere between curious and assisted right now. That's not a criticism. It's a map, and maps are how you move.
What this definition leaves out, on purpose
Nothing above mentions age, firm size, or any specific tool. A thirty-year partner can be AI-native, and a twenty-five year old who can prompt anything but reviews nothing can fail every mark. It also doesn't mention replacing people. Every mark above assumes a licensed human whose judgment got more valuable, not less.
Argue with me
This is a working definition and I fully expect to revise it. If you'd draw a line somewhere else, or you think a mark is missing, I genuinely want to hear it. The profession is going to end up with a definition of AI-native one way or another. I'd rather we write it than have it written for us.